Would like to find an inexpensive mini pc that would work good running umbrel and some of its apps, any good recommendations, thanks
I’m assuming cost is an important factor. Normally, I would suggest the GMKTec G3, but with RAM and SSD prices in crisis, the device itself that you install them into is becoming less and less important to the overall cost of the build.
I’m doing a “thinking outside-the-box” full node build which attempts to side-step these choke points (by properly configuring old-gen tech). When I finish that, I’ll test Umbrel on it and let you know what I find. Initial estimates are looking like it can be built for around $180 in the “USFF” form factor (i.e. tiny/mini PC)
Great, that being a good price point, what would the $300 range get me as far as a good setup. Would get me the Ram and SSD I
Would need for a good setup. Kinda want to use it for a server for AI contain within unbrel , I’m using an old Dell with intel core i5 pro 8th Gen.
Thanks
Tony
For AI, you need vRAM (unless you are using cloud models, in which case, the system specs are not terribly important). In the miniPC form factor, that really means a system with unified memory (since a standard GPU is bigger than the PC itself, haha)
Most people reach for the Mac Mini series, but they aren’t the only option for mini PCs with a unified memory architecture. The GMKtec EVO-X2 AI Mini PC, for example, can deliver similar specs for a lot lower price. The other option is refurbished Mac Minis. That said, the price for these systems ranges from $1,600 to $3,000+, depending on the amount of ram (64 GB on the lower end and 128 GB on the upper end)
For something in the $300 range, you are talking about using cloud AI models, in which case the system specs are going to be dominated by the requirements of the full node, not the AI work. This will be challenging if you want to run a full node and have a system that uses DDR4 (or DDR5) RAM and SSD storage. The minimum for RAM is 16 GB, and the minimum for storage is 2 TB. If you dial down the frequency of the RAM, the lowest price you can get 16 GB DDR4 is around $106 on Amazon, and the lowest price for 2TB SSD (if you go SATA instead of NVMe) is $190 (which would leave you only $4 for the device to install them into).
So I think the best bet for the $300 price range today (with the current RAM and SSD crisis) is to not go new on the RAM. Instead, get an older device that has 16GB DDR4 already installed in it, then buy it plus the 2TB SATA III SSD. One opition that i just built for another forum member which works for Start9 (and thus probably also Umbrel, since both are Debian under the hood) is the Lenovo ThinkCentre m710q Tiny, which you can usually find on eBay for around $150 - $170 (this one here, for example). Then you would get a cheap 2TB SATA SSD like this one here for around $190, which would put your total buld at around $360. Still over budget, but you might be able to shave off some of the price by bidding on the auctions rather than “Buy it now”.
If you are asking about my “thinking outside the box” build and investing a little more to improve on its specs, that one involves getting a device with DDR3 RAM (which has not been affected by the RAM crises currently), using a small SSD for the OS and a swap partition (to avoid file I/O on the data drive) and a large 2TB HDD for the data drive. Basically you install the OS on the SSD, add a 24GB+ swap partiton also on the SSD (resize the OS partition first if necessary to make room for it), and use the HDD for the data. The parts I am using in my proof of concept:
$40
16GB DDR3 RAM
$27
128GB mSATA SSD
$70
2TB 2.5" HDD
(plus taxes, charges, etc.)
This build maxes out at 16GB RAM, though, so no options to upgrade. If you wanted to go this route (and wanted a USFF/ tiny PC) then you’d have to research what options support more than 16GB DDR3 RAM (usually you need 4 RAM slots for that, which tend to come in the larger form factors)
I suppose you could go with a DDR4 option, but leveraging the small SSD + large HDD concept. Probably a variant on the Lenovo ThinkCentre m710q with 16GB RAM already installed could be worth exploring.
You mentioned building a setup for someone, is that something you do often, I want to run openclaw on a separate dedicated computer, would one the minisform pc perform will for that,
If you are using cloud LLM models, then any internet-connected computer should work fine (people even run OpenClaw on Raspberry Pi’s like this).
I recommend the Ollama cloud models, which have a good variety of open-source models (glm-5:cloud is what I use), new models added almost as soon as they get released, and all for fixed monthly billing (either $20 or $100) so you don’t get hit with surprise bills. OpenClaw is very token-intensive, so token-based billing is a budget disaster waiting to happen IMO.
Another good option if you have a ChatGPT Codex account already is the Codex OAuth method, which lets you use the plan you are already paying for. I think Nvidia also has a monthly billing plan for their cloud models as well.
What these miniPCs are not good for is running local models. For that you need vRAM. Cheapest option there are Mac Minis (which have a unified memory architecture, where system RAM is shared by both the CPU and GPU).
The other thing that you’d be limited on is some local tools that require vRAM (some of the memory architectures that do local Symantec searching, for example). Most AI tools can be done with cloud models (image generation, visual intelligence, voice interfaces, etc), you just have to do a little more work setting them up. Of course they also come with additional costs for credits and/or subscriptions.
Yes, I build systems for forum members frequently. I’ve done servers, air-gap systems, everyday PCs and laptops, and Bitcoin mining setups.
Great to know, have you done anything for a openclaw setup, from what I have read having it on a dedicate computer and maybe even in a docker container type environment is good , I’m not very techy but I been running it on umbrel on an older laptop, system resources are limiting it’s performance, so I wanted to try it on a separate machine, have done any with openclaw and how has your experience been, so far
You have given me some ideas and a previous message about some systems to look at. I was wondering if the minisform PC would be a good option in a specific configuration
Which Minisforum model were you thinking of using?
I’ve been running OpenClaw on a GMKTec G3 with 32GB RAM for a few weeks (using glm-5:cloud with the Ollama $100/mo. subscription). It works fine on that one. The main thing I struggled with is the OOB memory options are garbage, and the system is vulnerable as hell to prompt injection. The system kind of assumes you will be running Claude Sonnet, which is pretty resistent to prompt injection and supports a huge context window – but the cost to run it in that configuration would be crazy expensive.
I designed a system to install over the OOB system that I call the “Evergreen Toolkit”. I’m still working out some kinks, but I’ll have it up on Github soon for others to use. Basically, you would clone it and tell your OpenClaw agent to follow the instructions to adapt it to your system. It adds multi-user memory system with baked-in “theory of mind”, cross-session memories, memory backup, full search of all historic conversations, and injection of salient memories into conversations without the agent needing to search.
The toolkit also has some core “evergreen” programs (where the name comes from) that run nightly to iteratively self-improve on 4 critical domains: memory, system health, architecture, and prompt injection. Basically, each program cycle involves discovery (find what is available on the internet that might be useful), planning, implementation, and an introspection. The security program also has a “red team/blue team” concept – red team thinks from the perspective of a hacker, and tries to find vulnerabilities that it would use to attack the system. Then the blue team comes up with defenses against those attack vectors.
I’m not sure about which , minisform, are there some that think would be good options for this type of application out the box, I still figuring out how all this works, I do fine it very interesting, I’m just not a very techy person, so I’m looking at YouTube how to do stuff , and that always best way to learn because of there a lot of things done differently. It seems that running a model locally might be less costly in the long run as long as your machine can handle it I think, I am using open router to try different models but seem to have some issues or it maybe because I running it on my umbrel / container
For using cloud models, then the specific model of minPC doesn’t really matter. You’d just want something with at least 32GB of RAM (so you can at least run tiny local models for specific tasks like semantic memory searching)
For running a local model as your primary interface, then a system with unified memory is going to be necessary. Note that you still would probably want a cloud model that you have the main model leverage very sparingly for specific tasks that require better precision, such as writing complex code or troubleshooting problems that may arise, but you can keep the cost in that area very low by only using it sparingly (and in that case token-based billing just when you need it makes better sense than a fixed monthly cloud subscription cost).
Obviously, the more intelligent the model you run locally, the better your experience will be (and the less often you will need to turn to a cloud model to assist with something) The very upper end would be a M3 Ultra Mac Studio with 512GB unified memory, but those are crazy expensive (like $25,000 last I checked) A more reasonable target would be a system with 128GB unified memory, and go with an off-brand instead of Apple. The GMKtec EVO-X2 AI that I mentioned earlier is around $3,000 for the 128GB model last I checked. If you prefer minisforum, there is the MS-S1 Max AI for a similar price point. The drawaback with these two is it is a bit more technical to set up (since they use AMD chips instead of NVidia which is more widely supported). A simpler alternative would be the ASUS Ascent GX10, which sells for around $3,500 last I checked – most AI models would just work OOB on it since it uses Nvidia chips.
I suppose it depends on what is your budget, and how much you want to iterate and tweak your system to fit your needs, as opposed to setting it up and things just work OOB. The latter is actully a bit difficult (I would classify OpenClaw itself as experimental/beta software), so it is really a question of what degree of iterating/tweaking you can tolerate.
I just experimenting, and learning, my budget is nowhere near that price, even though it would nice not to have a top, I remember you help me with my start9 setup and gave my some good options in a not break the bank pricing. So looking for an out of the box minipc that work well so and can maybe be upgraded as I learn. I’m leaning towards API/OAth type service and maybe small local model maybe , but not sure what’s out there in a lower price range , that fit that example and use case,
Unfortunately, I don’t believe any of the unified memory architectures support RAM upgrades (the RAM on these systems tends to be soldered onto the board). There are “shared memory” architectures, but those tend to max out the RAM that you can allocate to the iGPU at 48GB or less (so if you upgraded to 128GB RAM, you wouldn’t be able to utilize most of it for AI models) Running capable AI models locally has a pretty high entry cost, sadly.
I think only the standard PC + GPU setup presents an upgrade path (but of course has a much larger footprint than a miniPC) As newer GPU models come out, you can often find the previous-gen models used for reasonable cost. For example, the NVidia Quandro M6000 gets you 24GB vRAM for around $200, or the Quandro RTX 8000 gets you 48GB vRAM for around $2,000) You could then swap out the GPU as newer models are released, and better previous-gen models become obsolete.
In any case, for a local model on lower-end devices, you’d probably want to use the glm-4.7-flash model from Ollama, which is a very capable model for its size (can do coding and agentic work), which will run on relatively low vRAM (24GB minimum for most useful tasks, but you get larger context window and/or more consecutive chats with more vRAM).
There seem to be a big jump in prices of the 2 devices you mentioned between the 24gb and the 48gb, is there any middle ground, , the 24gb would work with the Ollama model I’m thinking, having a cloud model incorporate should be beneficial in a setup
I’m planning to run a local-models only OpenClaw instance with the M6000 GPU. That should lead to some interesting findings, and I’ll share them. 24GB vRAM should work based on my back-of-the-napkin math, but it is cutting it very close to the bare minimum. There may be gotchas I haven’t considered, so I don’t recommend building a system for that just yet unless you are willing to take the risk. Otherwise, I recommend waiting until I can confirm that OpenClaw can function in such a setup.
Ok, looking forward to your review
Looks for me due to costs of devices I will have to use a cloud based system , I do have a subscription with ChatGPT using OAth services, hopefully lowing my cost for an entry level setup,
I have seen some devices with the Mac mini form factor but in a lower price range, think that should give me a starter setup in your opinion